update subset_transform (#218)

Author: helloyongyang

llmc/compression/quantization/awq.py

@@ -195,13 +195,15 @@ def block_transform(self, block, input_feat, block_kwargs):
     @torch.no_grad()
     def subset_transform(
         self,
-        layers_dict,
+        subset,
         input_feat,
-        prev_op,
-        input_name,
-        inspect_module,
         subset_kwargs,
     ):
+        layers_dict = subset['layers']
+        prev_op = subset['prev_op']
+        input_name = subset['input'][0]
+        inspect_module = subset['inspect']
+
         if not check_do_quant(
             self.block_idx,
             list(layers_dict.keys())[0],

llmc/compression/quantization/base_blockwise_quantization.py

@@ -465,10 +465,8 @@ def block_transform(self, block, input_feat, block_kwargs):
 
         for index, subset in enumerate(subsets):
             logger.info(f'subset: {subset}')
-            prev_op = subset['prev_op']
             layers_dict = subset['layers']
             input_name = subset['input'][0]
-            inspect_module = subset['inspect']
             inspect_has_kwargs = subset['has_kwargs']
             if inspect_has_kwargs:
                 if 'sub_keys' in subset:
@@ -478,11 +476,8 @@ def block_transform(self, block, input_feat, block_kwargs):
             else:
                 subset_kwargs = {}
             self.subset_transform(
-                layers_dict,
+                subset,
                 input_feat,
-                prev_op,
-                input_name,
-                inspect_module,
                 subset_kwargs,
             )
             if self.act_static:

llmc/compression/quantization/dgq.py

@@ -276,13 +276,14 @@ def search_scale_zero_subset(self, layers_dict, input_feat):
     @torch.no_grad()
     def subset_transform(
         self,
-        layers_dict,
+        subset,
         input_feat,
-        prev_op,
-        input_name,
-        inspect_module,
         subset_kwargs,
     ):
+        layers_dict = subset['layers']
+        prev_op = subset['prev_op']
+        input_name = subset['input'][0]
+
         layers = list(layers_dict.values())
         if isinstance(prev_op[0], tuple(_LLMC_LN_TYPES_ + _TRANSFORMERS_LN_TYPES_)):
             self.smoothquant_transform(prev_op, layers, input_feat[input_name])

llmc/compression/quantization/gptq.py

@@ -89,7 +89,13 @@ def block_transform(self, block, input_feat, block_kwargs):
         super().block_transform(block, input_feat, block_kwargs)
 
     @torch.no_grad()
-    def subset_transform(self, layers_dict, *args, **kwargs):
+    def subset_transform(
+        self,
+        subset,
+        input_feat,
+        subset_kwargs,
+    ):
+        layers_dict = subset['layers']
         for name in layers_dict:
             layer = layers_dict[name]
             self.layer_transform(layer, name)

llmc/compression/quantization/osplus.py

@@ -171,13 +171,15 @@ def search_scale_shift_subset(
     @torch.no_grad()
     def subset_transform(
         self,
-        layers_dict,
+        subset,
         input_feat,
-        prev_op,
-        input_name,
-        inspect_module,
         subset_kwargs,
     ):
+        layers_dict = subset['layers']
+        prev_op = subset['prev_op']
+        input_name = subset['input'][0]
+        inspect_module = subset['inspect']
+
         assert (
             len(prev_op) == 1
         ), 'Only support single prev_op. If multi prev_ops, code need to be updated.'

llmc/compression/quantization/rtn.py

@@ -19,11 +19,8 @@ def block_opt(self, *opt_kwargs):
     @torch.no_grad()
     def subset_transform(
         self,
-        layers_dict,
+        subset,
         input_feat,
-        prev_op,
-        input_name,
-        inspect_module,
         subset_kwargs,
     ):
         pass

llmc/compression/quantization/smoothquant.py

@@ -61,13 +61,14 @@ def search_scale_subset(self, layers, tensors):
     @torch.no_grad()
     def subset_transform(
         self,
-        layers_dict,
+        subset,
         input_feat,
-        prev_op,
-        input_name,
-        inspect_module,
         subset_kwargs,
     ):
+        layers_dict = subset['layers']
+        prev_op = subset['prev_op']
+        input_name = subset['input'][0]
+
         if not self.filter_subset(prev_op):
             logger.info('Do not transform this subset.')
             return

llmc/compression/sparsification/magnitude.py

@@ -14,13 +14,12 @@ def __init__(self, model, sparsity_config, input, padding_mask, config, modality
     @torch.no_grad()
     def subset_transform(
         self,
-        layers_dict,
+        subset,
         input_feat,
-        prev_op,
-        input_name,
-        inspect_module,
-        subset_kwargs
+        subset_kwargs,
     ):
+        layers_dict = subset['layers']
+
         layers = list(layers_dict.values())
         for layer in layers:
             W = layer.weight.data

llmc/compression/sparsification/wanda.py

@@ -33,13 +33,13 @@ def get_row_scale(self, layer, act):
     @torch.no_grad()
     def subset_transform(
         self,
-        layers_dict,
+        subset,
         input_feat,
-        prev_op,
-        input_name,
-        inspect_module,
-        subset_kwargs
+        subset_kwargs,
     ):
+        layers_dict = subset['layers']
+        input_name = subset['input'][0]
+
         layers = list(layers_dict.values())
         for layer in layers:
             scaler_row = self.get_row_scale(layer, input_feat[input_name][0])